BREAKING
AI Tools

Perplexity Computer Goes Fully Local on Mac: Setup, the 36GB Hardware Bar, and the Credit Catch

Updated October 22, 2026
9 min read
Editor's Note: This article was human-reviewed for accuracy
Open laptop glowing with neon light in a dark room, illustrating Perplexity Computer running as a local AI agent on a Mac

Quick Answer & Key Takeaways

  • • Perplexity's Computer agent now runs fully on your Mac — the model, the agent, and task execution all stay on-device, with data reaching the cloud only when you approve it.
  • • The catch is hardware: Perplexity recommends 36GB of unified memory, and it's available only to Pro, Max, and Enterprise subscribers.
  • • You still pay for cloud calls separately through Perplexity's credit system, where Pro gets no recurring monthly allowance.

The quick answer

  • What launched: Perplexity's Computer agent can now run fully locally on supported Apple silicon Macs — the model, the agent directing the task, and execution all stay on your machine (announced October 20, 2026).
  • What runs it: You choose between Qwen 3.6 and PPLX Pro, which Perplexity describes as its own post-trained version of Qwen. Planning, reasoning, and file processing happen on-device; code runs in isolated sandboxes.
  • The hardware bar: Perplexity recommends 36GB of unified memory or more. Most Macs Apple sells can't be configured with that much.
  • Who gets it: Pro ($20/month), Max ($200/month), and Enterprise subscribers only — the free tier doesn't get local mode, or Computer at all.
  • The pricing catch: Local mode gives you more control over credit spend, but cloud work still costs credits — and Pro has no recurring monthly credit allowance, just a one-time bonus that expires in 30 days.

What Perplexity Computer's local mode actually does

Perplexity Computer is the company's agentic system for the Mac: you describe a task, and it plans the steps, reasons through them, and carries them out. The October 20 update moves the whole stack onto your machine: the model that does the thinking, the agent that directs the task, and the execution itself.

In practice, that means Computer can work across local files and supported native apps without sending your data to a cloud model first. Perplexity's example: it scans your recent messages and emails, flags deadlines, and writes an action list in Notes. The apps it names: Notes, Messages, Mail, Keynote, Numbers, and Excel.

The trust mechanism here is permission, not just location. Computer gets access to local files and apps only with your approval, and code plus tool execution run in isolated sandbox environments — sealed-off workspaces with controlled access to your files and connected apps. Location is the headline, but the permission layer is what keeps an autonomous agent from rummaging through everything.

The 36GB bar — and who it leaves out

Local mode needs Apple silicon and, in Perplexity's words, 36GB of unified memory "or more" for the best experience. That is a serious machine. A base MacBook Air tops out at 24GB, which means every Air on sale is excluded from the recommended spec. But Perplexity is explicitly highlighting the new Mac mini machines as the target hardware, which tells you exactly who this is for: people already sitting on high-memory desktops, or willing to buy one.

This is worth dwelling on, because "run AI locally" sounds like a democratizing move, and the hardware requirement quietly narrows it back down. If your Mac doesn't clear the bar, Perplexity's answer is Hybrid Compute — more on that below — which works on Macs with as little as 24GB of memory, according to MacRumors' reporting on the announcement. (One outlet, NeoTeo, cites a 32GB recommendation for Hybrid instead, so treat the exact Hybrid floor as reporting variance, not gospel.)

How to set up local Computer on your Mac

The setup is a short path through the Mac app, not a DIY model install:

  1. Install or update the Perplexity app for Mac. Local mode ships in the current Mac app; you need a Pro, Max, or Enterprise subscription — local Computer is not available on the free tier.
  2. Download a local model. The app offers Qwen 3.6 or PPLX Pro, which Perplexity describes as a version of Qwen it has trained further. This is a one-off download; the model then lives on your machine.
  3. Choose the fully local option when you start a task. The app lets you pick where each task runs — fully local, Hybrid Compute, or the cloud path — so you decide per task, not once forever.

There is no published download size for the model, and Perplexity hasn't said how long the initial download takes, so budget disk space and patience accordingly.

Local vs Hybrid Compute: the actual difference

Perplexity now offers three ways to run Computer, and the differences are about where your data goes and what it costs you:

ModeWhat runs locallyWhat goes to the cloudMemory guidance
Fully localModel, agent, planning, reasoning, file processingOnly what you authorize: current information, browser use, cloud-connected apps — plus optional calls to 15+ frontier models for harder problems36GB or more recommended
Hybrid Compute (launched September 2026)Local model processes private information on-deviceCloud frontier models handle research and reasoningAs little as 24GB (MacRumors); 32GB cited elsewhere
Cloud ComputerNothing — the agent runs server-sideEverything, like any hosted AI serviceNo local requirement

Hybrid Compute is the compromise position: your private files are processed by the local model while the heavy reasoning happens on cloud frontier models. It arrived in September 2026, and it's the option for people who want local processing but don't own a 36GB Mac.

The pricing fine print — where the real catch lives

Local mode is not free to use. Perplexity's Computer runs on a credit system, and the details are where Pro subscribers should look twice:

  • Pro costs $20/month ($200/year). But here's the part people miss: Pro gets no recurring monthly credit allowance. Consumer Pro receives a one-time 4,000-credit bonus that expires after 30 days — after that, you buy credits as needed at $1 per 100 credits.
  • Max costs $200/month ($2,000/year) and includes 10,000 Computer credits per month plus a one-time 35,000-credit bonus. This is the tier the credit system is really built for.
  • Unused monthly credits don't roll over. Whatever you don't spend by month's end is gone. Max accounts also carry a monthly spending cap that defaults to $200 and can be raised to $5,000.
  • Perplexity hasn't said how many credits a local run consumes. Local mode gives you more control over credit spend, but the company hasn't published the credit cost of a local task, so you can't budget it in advance. That's an honest gap in the announcement.

For context on what tasks cost when the cloud is involved: Perplexity's own complexity bands put a light task (say, summarizing an earnings call) at 100–350 credits ($1–$3.50), a complex multi-step report at 350–950 credits ($3.50–$9.50), a heavy research-plus-code job at 875–2,275 credits ($8.75–$22.75), and a mega project with video or large-scale data collection at 2,400–9,800 credits ($24–$98). A single mega task can eat almost Max's entire monthly allowance.

So the honest math for a Pro user: $20/month gets you into the room, but sustained agentic work on the cloud path is pay-as-you-go, and the economics of local mode are partly unknowable until Perplexity publishes per-task credit costs.

What "fully local" doesn't mean

The announcement is exciting, but a few claims deserve deflating:

  • It's not offline-everything. One headline called it "even offline," and the local default can indeed work without an internet connection for on-device tasks. But anything that needs current information, a web browser, or a cloud-connected app goes to the cloud — and only with your authorization.
  • The local brain is Qwen 3.6, not a frontier model. For genuinely hard problems, the local agent can call on more than 15 larger cloud models. That fallback is the quiet re-entry of the cloud into your "fully local" task — useful, but worth knowing about when you're auditing where your data went.
  • "More control over credit spend" isn't the same as "cheap." Perplexity's own phrasing (reported by MacRumors) is that local mode gives users more control over credit spend — a choice-of-mode claim, not a cost claim. Until per-task credit pricing for local runs is published, "control" is all we can verify.
  • There's no Windows path. Local Computer is Apple silicon Macs only. The rest of the world waits.

Who this is actually for

Local Computer fits a specific buyer: a Pro or Max subscriber on a 36GB+ Apple silicon Mac who wants an agent inside their email, notes, and documents — and who has kept agents away from that data until now precisely because of where it would go. The industry pattern is worth naming: this is the same week Walmart opened its platform to shopping agents while Amazon blocked them, and Perplexity is pushing in the same local-first direction — agents that touch your real data are the new battleground, and the differentiator is where that data lives. If you want the contrast case, Walmart's agent gateway runs the opposite playbook: your data and the agent both live in the cloud, with spending limits as the guardrail. And if you want the full local-model extreme, Recursive's Flint runs a 40-billion-parameter coding model entirely on a laptop — no cloud at all.

Pros and cons

Pros

  • Real on-device privacy: model, agent, and task execution can all stay on your Mac
  • Granular control — you choose per task whether it runs locally, hybrid, or in the cloud
  • Permission-gated access to native Mac apps and files
  • Optional fallback to 15+ frontier models when the local model isn't enough

Cons

  • 36GB unified memory requirement excludes most Macs in Apple's lineup, including every MacBook Air
  • Pro subscribers get no recurring monthly credits — cloud calls beyond the one-time bonus are pay-as-you-go
  • No published credit cost for local runs, so budgeting is guesswork
  • Mac-only, Apple silicon only; no Windows or web version of local mode
  • The "fully local" framing breaks at the edges: web research, connected apps, and frontier-model calls still reach the cloud

FAQ

How do I run Perplexity Computer locally on my Mac?

Update the Perplexity Mac app, download one of the two local models (Qwen 3.6 or PPLX Pro) inside the app, then select the fully local option when starting a task. You need a Pro, Max, or Enterprise subscription and an Apple silicon Mac with 36GB of memory or more for the recommended experience.

How much RAM do I actually need?

Perplexity recommends 36GB of unified memory or more for fully local mode. Hybrid Compute — which keeps private processing local but sends reasoning to the cloud — works on Macs with as little as 24GB per MacRumors' reporting, though one outlet cites 32GB. Either way, a base-spec MacBook Air at 24GB sits at or below the recommended floor.

Is Perplexity Computer local mode really offline?

Mostly. The default keeps planning, reasoning, and file work on-device. But tasks that need current information, web browsing, or cloud-connected apps go to the cloud — only after you authorize them.

How much does Perplexity Pro cost, and what does it include for Computer?

Pro costs $20 per month ($200/year) and includes Computer access plus a one-time 4,000-credit bonus that expires in 30 days — no recurring monthly credits. Max costs $200/month ($2,000/year) and includes 10,000 Computer credits per month plus a 35,000-credit one-time bonus. Credits convert at $1 per 100, and unused monthly credits don't roll over.

Does local Computer work on Windows or in the browser?

No. The fully local mode is available only in the Perplexity app for Mac, on supported Apple silicon Macs. There's no Windows version and no browser-based local mode.

Key takeaways

  • Perplexity Computer's fully local mode, announced October 20, 2026, runs the model, the agent, and task execution on your Mac — cloud access requires your explicit approval per task.
  • The 36GB unified-memory recommendation means most Macs can't run it; Hybrid Compute at 24GB is Perplexity's answer for everyone else.
  • Pro ($20/mo) includes Computer but no recurring credits — just a one-time 4,000-credit bonus expiring in 30 days — while Max ($200/mo) bundles 10,000 monthly credits.
  • "Fully local" has real edges: web research, connected apps, and frontier-model fallbacks still reach the cloud, and Perplexity hasn't published credit costs for local runs.
  • It's Mac-only and subscription-only — but it's the clearest signal yet that local-first agents are the industry's next differentiator.

Local-first agents are no longer a research demo — as of October 22, 2026, you can buy one in the Mac App Store's neighborhood, provided your Mac has the memory to run it.

9 min read
5 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the Founder and Editor-in-Chief of AI Core Updates, steering the publication's strategic growth, technological architecture, and human-verified journalistic standards. With over eight years of hands-on experience across full-stack systems and machine learning infrastructure, Ratul founded AI Core Updates to deliver transparent, benchmark-backed intelligence that bridges the divide between foundational research and real-world executive decision-making. At AI Core Updates, he directs long-term editorial investigations and leads the development of internal automated verification workflows.

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

AI Core Updates Insider Dispatch

Stay Ahead of the Exponential AI Curve

Join over 48,000+ AI researchers, software engineers, and founders. Get our daily 5-minute morning intelligence briefing analyzing breakthrough research, frontier models, and autonomous workflows.

bolt
Daily Briefing in 5 Mins
High signal, zero marketing fluff.
podcasts
Podcast Deep Dives
Early access & episode transcripts.
psychology
Evaluated Prompts & Tools
Tested benchmarks & system designs.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
AK
SL
MR
JD
Trusted by builders from leading AI labs & enterprises.
mark_email_unread

Subscribe to Newsletter

Delivered daily at 7:00 AM UTC

person_outline
mail_outline
Include in my subscription:

No spam ever. You can easily unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more